Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A computer-implemented method for determining routing, comprising: inputting a plurality of to-be-optimized routing solution candidates to a Siamese neural network comprising a plurality of value prediction networks, wherein: each of the value prediction networks being trained to predict a cost associated with a to-be-optimized routing solution candidate, and each of the plurality of to-be-optimized routing solution candidates comprises one or more routes for routing one or more vehicles through a plurality of locations; identifying one or more to-be-optimized routing solution candidates from the plurality of to-be-optimized routing solution candidates based on outputs of the Siamese neural network; inputting the one or more identified to-be-optimized routing solution candidates to a routing optimizer to obtain one or more optimized routing solution candidates, wherein the routing optimizer comprises a set of improvement operators performing one or more of following operations: changing an order of at least two of the plurality of locations in one of the one or more routes, and moving a location from one of the one or more routes to another one of the one or more routes; and determining an optimized routing solution with a lowest cost from the one or more optimized routing solution candidates.
This invention relates to computer-implemented methods for optimizing vehicle routing solutions. The problem addressed is efficiently finding the best routes for vehicles to travel through multiple locations, minimizing associated costs. The method involves using a Siamese neural network to evaluate multiple potential routing solutions. This network contains several value prediction networks, each trained to estimate the cost of a given routing solution candidate. A routing solution candidate can include one or more routes for one or more vehicles, specifying the sequence of locations to be visited. Based on the cost predictions from the Siamese neural network, one or more promising routing solution candidates are selected. These selected candidates are then fed into a routing optimizer. The optimizer refines these solutions using improvement operators. These operators can alter the order of locations within a route or relocate a location from one route to another. Finally, the method determines the overall optimized routing solution by selecting the candidate with the lowest cost from the set of refined solutions produced by the routing optimizer.
2. The method of claim 1 , wherein each of the plurality of to-be-optimized routing solution candidates is subject to one or more constraints, the constraints comprising one or more of the following: time constraint; travel distance constraint; vehicle capacity constraint; and power expense constraint.
This invention relates to optimizing routing solutions for transportation or logistics systems, particularly where multiple constraints must be balanced. The method generates a plurality of routing solution candidates and evaluates each candidate against one or more constraints to determine an optimal solution. The constraints include time constraints, such as minimizing travel time or ensuring delivery within a specified window; travel distance constraints, such as minimizing total distance or adhering to route length limits; vehicle capacity constraints, such as ensuring cargo does not exceed vehicle limits; and power expense constraints, such as minimizing fuel or energy consumption. The method may apply these constraints individually or in combination to refine the routing solutions. The goal is to improve efficiency, cost-effectiveness, and feasibility in transportation planning by systematically evaluating and optimizing routes under real-world operational limitations. This approach is useful in logistics, fleet management, and delivery services where multiple competing factors must be balanced to achieve practical and efficient routing solutions.
3. The method of claim 2 , wherein the set of improvement operators in the routing optimizer are learned based on a reinforcement learning algorithm.
A routing optimization system improves the efficiency of delivery or logistics networks by dynamically adjusting routes in real-time. The system addresses the problem of static routing plans that fail to adapt to changing conditions such as traffic, weather, or vehicle availability, leading to inefficiencies in fuel consumption, delivery times, and resource utilization. The routing optimizer uses a set of improvement operators to refine routes. These operators are not predefined but are instead learned through a reinforcement learning algorithm. The algorithm evaluates the performance of different routing adjustments in simulated or real-world scenarios, reinforcing operators that lead to better outcomes such as shorter travel times or lower costs. Over time, the system improves its ability to generate optimal routes by continuously updating the operators based on feedback from past decisions. The reinforcement learning approach allows the system to adapt to new conditions without manual intervention, making it more flexible and scalable than traditional rule-based or heuristic methods. This self-learning capability ensures that the routing optimizer remains effective even as operational constraints or environmental factors change. The system can be applied to various logistics applications, including last-mile delivery, fleet management, and supply chain optimization.
4. The method of claim 1 , wherein the plurality of value prediction networks in the Siamese neural network are identical.
A system and method for value prediction using a Siamese neural network architecture addresses the challenge of accurately predicting values in scenarios where data relationships are complex or non-linear. The invention employs a neural network with multiple value prediction networks, each designed to process input data and generate predicted values. The key innovation lies in the use of identical value prediction networks within the Siamese structure, ensuring consistency and symmetry in the learning process. This approach enhances the network's ability to capture intricate patterns and dependencies in the data, improving prediction accuracy. The identical networks share weights, allowing the system to leverage shared knowledge across different parts of the input space. This method is particularly useful in applications such as financial forecasting, risk assessment, and recommendation systems, where precise value predictions are critical. The identical structure of the prediction networks simplifies training and reduces computational overhead while maintaining high performance. By standardizing the network architecture, the system ensures robustness and reliability in diverse real-world applications.
5. The method of claim 1 , wherein the Siamese neural network comprises two value prediction networks, and the method further comprises: training the two value prediction networks by performing one or more iterations of a tuning process, wherein the performing one or more iterations of a tuning process comprises: obtaining a training set comprising a third to-be-optimized routing solution candidate and a fourth to-be-optimized routing solution candidate; inputting the training set to the routing optimizer to obtain a third score for the third to-be-optimized routing solution candidate and a fourth score for the fourth to-be-optimized routing solution candidate; inputting the training set to the two value prediction networks respectively to obtain a fifth score for the third to-be-optimized routing solution candidate and a sixth score for the fourth to-be-optimized routing solution candidate; and tuning weights of the two value prediction networks based at least on the third score, the fourth score, the fifth score, and the sixth score.
The invention relates to optimizing routing solutions using a Siamese neural network with two value prediction networks. The technology addresses the challenge of efficiently evaluating and improving routing solutions in optimization tasks, such as logistics or network routing, where multiple candidate solutions must be compared and refined. The method involves training the two value prediction networks through an iterative tuning process. During each iteration, a training set containing two routing solution candidates is provided. The routing optimizer evaluates these candidates, generating scores for each. The same candidates are then input into the two value prediction networks, which also produce scores. The weights of the value prediction networks are adjusted based on the scores from both the routing optimizer and the prediction networks. This tuning process ensures that the value prediction networks align their evaluations with the routing optimizer's assessments, improving their accuracy in scoring future routing solutions. The Siamese neural network structure allows the two value prediction networks to learn from each other, enhancing their ability to distinguish between high-quality and low-quality routing solutions. This approach improves the efficiency and effectiveness of routing optimization by reducing reliance on computationally expensive evaluations from the routing optimizer.
6. The method of claim 5 , wherein prior to the inputting the training set to the two value prediction networks, the tuning process further comprises: determining whether a difference between the third score and the fourth score is greater than a preset threshold; and if not, abandoning the third to-be-optimized routing solution candidate and the fourth to-be-optimized routing solution candidate.
This invention relates to optimizing routing solutions in a machine learning system, specifically for improving the performance of value prediction networks used in routing optimization tasks. The problem addressed is the inefficiency in evaluating and selecting optimal routing solutions when multiple candidate solutions yield similar performance scores, leading to unnecessary computational overhead. The method involves a tuning process for two value prediction networks that evaluate routing solution candidates. Before inputting a training set to these networks, the system first compares the performance scores (third and fourth scores) of two routing solution candidates (third and fourth candidates). If the difference between these scores is not greater than a preset threshold, indicating that the candidates perform similarly, the system abandons both candidates to avoid further unnecessary processing. This selective filtering ensures that only significantly distinct candidates are retained for further optimization, improving computational efficiency. The tuning process also includes evaluating the performance of the value prediction networks themselves, ensuring that the networks are effectively distinguishing between high and low-quality routing solutions. By dynamically adjusting the threshold, the system can balance between computational efficiency and solution quality. This approach is particularly useful in large-scale routing optimization tasks where evaluating all possible candidates is impractical.
7. The method of claim 5 , wherein the tuning weights of the two value prediction networks comprises: determining a label for the training set based on the third score and the fourth score; converting the fifth score and the sixth score to a fifth logit value and a sixth logit value; determining a cross-entropy loss function based on the label, the fifth logit value and the sixth logit value; and tuning weights of the two value prediction networks based on the cross-entropy loss function.
This invention relates to machine learning systems for value prediction, specifically improving the training of multiple value prediction networks. The problem addressed is optimizing the tuning of weights in these networks to enhance prediction accuracy. The method involves using a cross-entropy loss function derived from logit values and labels to adjust the weights of two value prediction networks. The process begins by determining a label for the training set based on two scores generated by the networks. These scores are then converted into logit values. A cross-entropy loss function is computed using the label and the logit values. This loss function quantifies the discrepancy between the predicted and actual values, guiding the weight tuning process. The weights of the two value prediction networks are then adjusted based on the computed loss to minimize prediction errors and improve performance. The method ensures that the networks learn from their predictions by leveraging cross-entropy loss, a common technique in machine learning for classification tasks. By iteratively refining the weights, the networks become more accurate in their value predictions. This approach is particularly useful in applications requiring precise value estimation, such as reinforcement learning or decision-making systems. The technique enhances the robustness and reliability of the prediction models.
8. The method of claim 4 , wherein each of the plurality of value prediction networks comprises: a bidirectional Long Short-Term Memory (LSTM) layer comprising a plurality of LSTM units; an attention layer for embedding outputs from the plurality of LSTM units; and an output layer for generating a score based on an output from the attention layer and a plurality of features associated with a to-be-optimized routing solution candidate.
This invention relates to a machine learning-based system for optimizing routing solutions, particularly in scenarios where multiple candidate routes must be evaluated and scored. The problem addressed is the need for accurate and efficient prediction of route performance metrics, such as delivery time or cost, to select the best routing solution from a set of candidates. The system employs a plurality of value prediction networks, each designed to assess different aspects of a routing solution. Each network includes a bidirectional Long Short-Term Memory (LSTM) layer, which processes sequential data in both forward and backward directions to capture temporal dependencies in the routing data. The LSTM layer consists of multiple LSTM units that extract relevant features from the input sequence. An attention layer then embeds the outputs from the LSTM units, allowing the network to focus on the most relevant parts of the sequence for predicting the route's value. This mechanism improves the accuracy of the predictions by dynamically weighting the importance of different time steps in the sequence. Finally, an output layer generates a score for the routing solution candidate based on the attention layer's output and additional features associated with the candidate. These features may include factors like distance, traffic conditions, or vehicle capacity, which influence the route's performance. By combining these components, the system provides a robust framework for evaluating and optimizing routing solutions in real-world applications, such as logistics and transportation.
9. The method of claim 8 , wherein each of the one or more routes in the to-be-optimized routing solution candidate is associated with a distance, and the plurality of features associated with the to-be-optimized routing solution candidate comprise: a sum of the distances of the one or more routes in the to-be-optimized routing solution candidate; and a standard deviation of the distances of the one or more routes in the to-be-optimized routing solution candidate.
The invention relates to optimizing routing solutions in logistics or transportation systems, particularly for scenarios where multiple routes must be balanced for efficiency. The problem addressed is the need to evaluate and optimize routing solutions not just by total distance but also by route variability, ensuring fairness and efficiency across different routes. The method involves analyzing a candidate routing solution that includes one or more routes, where each route has an associated distance. The solution candidate is evaluated based on two key features: the sum of the distances of all routes in the candidate solution, representing total travel distance, and the standard deviation of those distances, representing route variability. By considering both metrics, the method enables optimization that balances minimizing total distance while also reducing disparities between individual route lengths, leading to more equitable and efficient routing solutions. This approach is useful in applications like delivery logistics, fleet management, or any system requiring multi-route optimization with fairness constraints.
10. The method of claim 1 , wherein the Siamese neural network comprises two value prediction networks, and the inputting a plurality of to-be-optimized routing solution candidates to a Siamese neural network comprises: for each to-be-optimized routing solution candidate of the plurality of the to-be-optimized routing solution candidates: pairing the to-be-optimized routing solution candidate with each other to-be-optimized routing solution candidate of the plurality of to-be-optimized routing solution candidates that is different from the to-be-optimized routing solution candidate; and inputting the to-be-optimized routing solution candidate and the paired each other candidate into the two value prediction networks to determine an individual score of the to-be-optimized routing solution candidate.
This invention relates to optimizing routing solutions using a Siamese neural network architecture. The problem addressed is efficiently evaluating and comparing multiple routing solution candidates to identify the most optimal one. Traditional methods often struggle with scalability and accuracy when assessing large sets of potential routes. The solution involves a Siamese neural network composed of two value prediction networks. Each routing solution candidate is paired with every other distinct candidate in the set. These pairs are then input into the two value prediction networks, which independently evaluate each candidate. The network generates an individual score for each routing solution based on its performance relative to the paired candidate. This pairwise comparison approach allows the system to learn and rank routing solutions more effectively by leveraging the shared structure of the Siamese network, which ensures consistent evaluation criteria across all candidates. The method improves optimization by providing a robust, data-driven way to assess and select the best routing solution from a large pool of possibilities.
11. The method of claim 10 , wherein the inputting the to-be-optimized routing solution candidate and the paired each other to-be-optimized routing solution candidate into the two value prediction networks to determine an individual score of the to-be-optimized routing solution candidate comprises: obtaining a seventh score for the to-be-optimized routing solution candidate and an eighth score for the paired each other to-be-optimized routing solution candidate; determining a positive score for the to-be-optimized routing solution candidate if the seventh score is greater than the eighth score; and determining a non-positive score for the to-be-optimized routing solution candidate if the seventh score is not greater than the eighth score.
This invention relates to optimizing routing solutions, particularly in systems where multiple routing options must be evaluated and compared. The problem addressed is the need for an efficient and accurate method to assess the quality of routing solutions by comparing them against paired alternatives. The method involves using two value prediction networks to evaluate routing solution candidates. Each candidate is paired with another candidate, and both are input into the networks to generate individual scores. The first network produces a seventh score for the primary candidate, while the second network generates an eighth score for the paired candidate. The primary candidate receives a positive score if its seventh score exceeds the eighth score, indicating it is superior. Conversely, if the seventh score is not greater, the primary candidate receives a non-positive score, indicating it is not better than the paired candidate. This approach enables a comparative evaluation of routing solutions, helping to identify the most optimal option based on predicted performance metrics. The method is particularly useful in applications requiring real-time or near-real-time routing optimization, such as logistics, transportation, or network management.
12. The method of claim 10 , wherein the identifying one or more to-be-optimized routing solution candidates from the plurality of to-be-optimized routing solution candidates based on the Siamese neural network comprises: for each to-be-optimized routing solution candidate of the plurality of the to-be-optimized routing solution candidates: determining an overall score for the to-be-optimized routing solution candidate comprising a sum of the individual scores of the to-be-optimized routing solution candidate; and identifying a to-be-optimized routing solution candidate from the plurality of to-be-optimized routing solution candidates with a highest overall score.
This invention relates to optimizing routing solutions using a Siamese neural network, particularly in scenarios where multiple routing options must be evaluated and prioritized. The problem addressed is the need for an efficient and accurate method to identify the most optimal routing solution from a set of candidates, considering various performance metrics. The method involves evaluating a plurality of routing solution candidates by determining an overall score for each candidate. This overall score is calculated as the sum of individual scores assigned to the candidate based on specific criteria. A Siamese neural network is used to generate these individual scores by comparing the candidate routing solutions against reference or benchmark solutions. The network processes each candidate solution to produce a similarity or performance score, which is then aggregated into an overall score. The routing solution candidate with the highest overall score is identified as the most optimized option. This approach leverages the Siamese neural network's ability to learn and compare complex patterns in routing data, ensuring that the selected solution meets predefined optimization criteria. The method is particularly useful in applications requiring real-time or near-real-time routing decisions, such as logistics, transportation, or network management.
13. A system for determining routing, comprising one or more processors and one or more non-transitory computer-readable memories coupled to the one or more processors and configured with instructions executable by the one or more processors to cause the system to perform operations comprising: inputting a plurality of to-be-optimized routing solution candidates to a Siamese neural network comprising a plurality of value prediction networks, wherein: each of the value prediction networks being trained to predict a cost associated with a to-be-optimized routing solution candidate, and each of the plurality of to-be-optimized routing solution candidates comprises one or more routes for routing one or more vehicles through a plurality of locations; identifying one or more to-be-optimized routing solution candidates from the plurality of to-be-optimized routing solution candidates based on outputs of the Siamese neural network; inputting the one or more identified to-be-optimized routing solution candidates to a routing optimizer to obtain one or more optimized routing solution candidates, wherein the routing optimizer comprises a set of improvement operators performing one or more of following operations: changing an order of at least two of the plurality of locations in one of the one or more routes, and moving a location from one of the one or more routes to another one of the one or more routes; and determining an optimized routing solution with a lowest cost from the one or more optimized routing solution candidates.
The system optimizes routing solutions for vehicles navigating multiple locations by leveraging machine learning and optimization techniques. The problem addressed is the computational complexity of evaluating and refining routing solutions, particularly in scenarios with numerous possible routes and constraints. The system uses a Siamese neural network with multiple value prediction networks to assess the cost or efficiency of routing solution candidates. Each candidate consists of one or more routes for vehicles, covering multiple locations. The neural network identifies promising candidates based on predicted costs, reducing the search space for further optimization. These selected candidates are then processed by a routing optimizer, which applies improvement operators to refine the routes. These operators can reorder locations within a route or transfer locations between routes to enhance efficiency. The system ultimately selects the optimized routing solution with the lowest cost from the refined candidates. This approach combines predictive modeling with heuristic optimization to improve routing efficiency in logistics, transportation, or delivery applications.
14. The system of claim 13 , wherein the Siamese neural network comprises two value prediction networks, and the operations further comprise: training the two value prediction networks by performing one or more iterations of a tuning process, wherein the performing one or more iterations of a tuning process comprises: obtaining a training set comprising a third to-be-optimized routing solution candidate and a fourth to-be-optimized routing solution candidate; inputting the training set to the routing optimizer to obtain a third score for the third to-be-optimized routing solution candidate and a fourth score for the fourth to-be-optimized routing solution candidate; inputting the training set to the two value prediction networks respectively to obtain a fifth score for the third to-be-optimized routing solution candidate and a sixth score for the fourth to-be-optimized routing solution candidate; and tuning weights of the two value prediction networks based at least on the third score, the fourth score, the fifth score, and the sixth score.
The system relates to optimizing routing solutions using a Siamese neural network with two value prediction networks. The technology addresses the challenge of efficiently evaluating and improving routing solutions in complex systems, such as logistics or network optimization, where multiple candidate solutions must be compared and refined. The system includes a routing optimizer that generates scores for routing solution candidates. The Siamese neural network consists of two value prediction networks, each trained to assess the quality of routing solutions. During training, the system iteratively refines these networks by comparing their predictions against the routing optimizer's scores. A training set containing two routing solution candidates is input into the routing optimizer, producing scores for each. The same candidates are then evaluated by the two value prediction networks, generating additional scores. The weights of the value prediction networks are adjusted based on the differences between the routing optimizer's scores and the networks' predictions, improving their accuracy over time. This process ensures the networks learn to align their evaluations with the routing optimizer's assessments, enhancing the overall optimization capability.
15. The system of claim 14 , wherein prior to the inputting the training set to the two value prediction networks, the tuning process further comprises: determining whether a difference between the third score and the fourth score is greater than a preset threshold; and if not, abandoning the third to-be-optimized routing solution candidate and the fourth to-be-optimized routing solution candidate.
This invention relates to optimizing routing solutions in a system that uses machine learning models, specifically two value prediction networks, to evaluate and refine routing options. The problem addressed is the computational inefficiency and suboptimal performance that can arise when evaluating routing solutions that are too similar, leading to redundant processing and wasted resources. The system includes a tuning process that compares the performance scores of two routing solution candidates. Before inputting a training set to the two value prediction networks, the system first determines whether the difference between the third score (associated with one routing solution candidate) and the fourth score (associated with another routing solution candidate) exceeds a preset threshold. If the difference is not greater than the threshold, indicating that the routing solutions are too similar, the system abandons both candidates to avoid unnecessary processing. This selective filtering ensures that only meaningfully distinct routing solutions are further evaluated, improving efficiency and computational performance. The value prediction networks are then used to refine the remaining routing solutions based on the training set, ensuring optimal routing decisions. The invention enhances the efficiency of routing optimization by eliminating redundant candidates early in the process.
16. The system of claim 14 , wherein the Siamese neural network comprises two value prediction networks, and the inputting a plurality of to-be-optimized routing solution candidates to a Siamese neural network comprises: for each to-be-optimized routing solution candidate of the plurality of the to-be-optimized routing solution candidates: pairing the to-be-optimized routing solution candidate with each other to-be-optimized routing solution candidate of the plurality of to-be-optimized routing solution candidates that is different from the to-be-optimized routing solution candidate; and inputting the to-be-optimized routing solution candidate and the paired each other candidate into the two value prediction networks to determine an individual score of the to-be-optimized routing solution candidate.
The system relates to optimizing routing solutions using a Siamese neural network. The problem addressed is efficiently evaluating multiple routing solution candidates to identify the most optimal one. Traditional methods often struggle with scalability and accuracy when comparing numerous routing options. The system includes a Siamese neural network with two value prediction networks. For each routing solution candidate, the system pairs it with every other distinct candidate. Each pair is then input into the two value prediction networks, which generate an individual score for the candidate. This pairwise comparison allows the network to learn similarities and differences between solutions, enabling more accurate optimization. The scores are used to rank or select the best routing solution based on predefined criteria, such as efficiency, cost, or time. The approach improves decision-making in routing optimization by leveraging deep learning to handle complex comparisons efficiently.
17. The system of claim 16 , wherein the identifying one or more to-be-optimized routing solution candidates from the plurality of to-be-optimized routing solution candidates based on the Siamese neural network comprises: for each to-be-optimized routing solution candidate of the plurality of the to-be-optimized routing solution candidates: determining an overall score for the to-be-optimized routing solution candidate comprising a sum of the individual scores of the to-be-optimized routing solution candidate; and identifying a to-be-optimized routing solution candidate from the plurality of to-be-optimized routing solution candidates with a highest overall score.
The invention relates to optimizing routing solutions in a system, particularly for selecting the most efficient routing paths from multiple candidates. The problem addressed is the need for an automated and accurate method to evaluate and prioritize routing solutions based on performance metrics. The system uses a Siamese neural network to assess routing solution candidates by comparing their individual scores, which are derived from factors such as distance, time, cost, or other relevant parameters. For each candidate, the system calculates an overall score by summing these individual scores. The candidate with the highest overall score is then identified as the optimal routing solution. This approach ensures that the best-performing route is selected based on a comprehensive evaluation of multiple factors, improving efficiency and reliability in routing decisions. The system is designed to handle large datasets and complex routing scenarios, making it suitable for applications in logistics, transportation, and network optimization.
18. A non-transitory computer-readable storage medium for determining routine, configured with instructions executable by one or more processors to cause the one or more processors to perform operations comprising: inputting a plurality of to-be-optimized routing solution candidates to a Siamese neural network comprising a plurality of value prediction networks, wherein: each of the value prediction networks being trained to predict a cost associated with a to-be-optimized routing solution candidate, and each of the plurality of to-be-optimized routing solution candidates comprises one or more routes for routing one or more vehicles through a plurality of locations; identifying one or more to-be-optimized routing solution candidates from the plurality of to-be-optimized routing solution candidates based on outputs of the Siamese neural network; inputting the one or more identified to-be-optimized routing solution candidates to a routing optimizer to obtain one or more optimized routing solution candidates, wherein the routing optimizer comprises a set of improvement operators performing one or more of following operations: changing an order of at least two of the plurality of locations in one of the one or more routes, and moving a location from one of the one or more routes to another one of the one or more routes; and determining an optimized routing solution with a lowest cost from the one or more optimized routing solution candidates.
This invention relates to optimizing routing solutions for vehicle fleets using machine learning. The problem addressed is efficiently identifying and refining high-quality routing solutions from a large set of candidates to minimize costs such as time, distance, or fuel consumption. The system uses a Siamese neural network with multiple value prediction networks, each trained to evaluate the cost of a given routing solution. Each routing solution candidate includes one or more routes for vehicles traveling through multiple locations. The Siamese network ranks or filters these candidates based on predicted costs. The top candidates are then processed by a routing optimizer, which applies improvement operators to refine the routes. These operators can reorder locations within a route or transfer locations between routes. The optimizer generates optimized routing solutions, and the system selects the one with the lowest cost. This approach combines machine learning for initial candidate evaluation with traditional optimization techniques for fine-tuning, improving efficiency in logistics and transportation planning.
19. The storage medium of claim 18 , wherein the Siamese neural network comprises two value prediction networks, and the operations further comprise: training the two value prediction networks by performing one or more iterations of a tuning process, wherein the performing one or more iterations of a tuning process comprises: obtaining a training set comprising a third to-be-optimized routing solution candidate and a fourth to-be-optimized routing solution candidate; inputting the training set to the routing optimizer to obtain a third score for the third to-be-optimized routing solution candidate and a fourth score for the fourth to-be-optimized routing solution candidate; inputting the training set to the two value prediction networks respectively to obtain a fifth score for the third to-be-optimized routing solution candidate and a sixth score for the fourth to-be-optimized routing solution candidate; and tuning weights of the two value prediction networks based at least on the third score, the fourth score, the fifth score, and the sixth score.
The invention relates to optimizing routing solutions using a Siamese neural network with two value prediction networks. The technology addresses the challenge of efficiently evaluating and improving routing solutions in complex systems, such as logistics or network optimization, where multiple candidate solutions must be compared and refined. The system includes a routing optimizer that generates scores for routing solution candidates and a Siamese neural network with two value prediction networks. The neural network is trained through an iterative tuning process. During training, a training set containing two routing solution candidates is input into the routing optimizer, which produces scores for each candidate. The same training set is then input into the two value prediction networks, which also generate scores for the candidates. The weights of the value prediction networks are adjusted based on the scores from both the routing optimizer and the value prediction networks. This tuning process is repeated to improve the accuracy and performance of the value prediction networks in evaluating routing solutions. The approach enhances the ability to compare and select optimal routing solutions efficiently.
20. The storage medium of claim 19 , wherein prior to the inputting the training set to the two value prediction networks, the tuning process further comprises: determine whether a difference between the third score and the fourth score is greater than a preset threshold; and if not, abandoning the third to-be-optimized routing solution candidate and the fourth to-be-optimized routing solution candidate.
This invention relates to optimizing routing solutions in a machine learning system, specifically for improving the performance of routing algorithms by evaluating and refining candidate solutions. The problem addressed is the inefficiency in selecting optimal routing paths due to suboptimal evaluation metrics, leading to poor performance in routing applications. The system uses two value prediction networks to evaluate routing solution candidates. Each network generates a score for a candidate solution, and these scores are compared. If the difference between the scores from the two networks is below a preset threshold, the candidate solutions are deemed insufficiently distinct and are discarded. This filtering step ensures that only meaningful and divergent candidates proceed to further optimization, improving computational efficiency and solution quality. The tuning process involves iterative evaluation of multiple routing solution candidates, where each candidate is assessed by the two prediction networks. By comparing the scores, the system identifies and eliminates candidates that do not meet the threshold for meaningful differentiation, thereby refining the pool of candidates for subsequent optimization steps. This approach enhances the reliability and effectiveness of the routing solution selection process.
Unknown
October 20, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.